AI Finance

AI Intention Economy

UK study reveals risks of AI intention economy, emphasizing ethics, transparency, and regulation for equitable growth.
Diagram showing the AI intention economy with chat prompts feeding into real-time intent auctions and buyer bids

Introduction

A new UK Study Highlights AI Intention Economy Risks and warns that chatbots may soon sell what you want before you say it. Cambridge researchers Yaqub Chaudhary and Jonnie Penn published their analysis in the Harvard Data Science Review on 30 December 2024. They argue that large language models can already infer motives from the words a user types into a chat window every day. Their paper describes a market where advertisers bid for a user’s future action, not the user’s attention or screen time. Independent forecasts suggest the commercial pull is already here, with eMarketer estimating that AI-powered ad spend will reach fifty-seven billion dollars in 2026. This article unpacks the study, the mechanisms behind the risk, the regulatory response in Europe, and what product teams can do today.

Quick Answers on the AI Intention Economy

What is the AI intention economy in simple terms?

The AI intention economy is a market where models infer, record, and sell predictions about what users plan to do next. Cambridge researchers say chatbots can turn every private conversation into a bidding signal.

Who wrote the UK intention economy study?

Yaqub Chaudhary and Jonnie Penn of the Cambridge Leverhulme Centre published the AI intention economy paper in the Harvard Data Science Review on 30 December 2024. Both authors specialise in the political economy of AI.

Is the AI intention economy already operating?

Pieces of the AI intention economy are operating, with agent tools like OpenAI Operator booking purchases on behalf of users. Full intent auctions are not yet live at scale, though the ingredients and incentives exist today.

Key Takeaways From the UK Intention Economy Warning

  • The Cambridge study says conversational AI can infer user intent from chat history, tone, and personal profile data without explicit consent.
  • Chaudhary and Penn frame intent as a new commodity that may trade on real-time auction markets, not simply as a product feature.
  • The researchers single out elections, a free press, and fair market competition as systems most exposed to intent-level manipulation.
  • Independent regulators, including the European Commission, already ban manipulative AI under Article 5 of the EU AI Act with fines up to seven percent of global turnover.

What Is the AI Intention Economy

The AI intention economy is a predicted market where models infer a user’s future actions, package those predictions as structured signals, and sell them to advertisers, platforms, or political actors through real-time auctions and agent intermediaries.

An Interactive From AIplusInfo

The Intent Commodification Calculator

Explore how an AI chatbot could price your intent across a typical session, using assumptions from the Cambridge Leverhulme paper and 2026 ad market data.


8

140

Moderate urgency

lowpremium

$42

$5$120

Annual intent value per user

$-

Based on your assumptions

Ad exposure equivalent
–
Share of 2026 US AI ad market ($57B)
–

CPM range based on eMarketer 2026 AI ad spend forecast and the Cambridge Leverhulme Centre paper on the intention economy.

Inside the Cambridge Leverhulme Study on Intent Commodification

Shifting focus to the paper itself, Chaudhary and Penn frame the AI intention economy as an emerging layer sitting on top of today’s attention-driven internet. Their central claim is that large language models can elicit, infer, collect, record, and sell signals of human intent at commercial scale. The authors draw on technology history, political economy, and recent product announcements to argue that the plumbing already exists. They anchor the warning in the Leverhulme Centre’s broader research on how to make AI systems accountable to the public, not only to shareholders. The paper was published in the open-access Harvard Data Science Review, so any product manager or policy analyst can read the full text without a paywall.

Chaudhary and Penn then detail how a chatbot could turn an ordinary conversation into a profile rich enough to drive bidding. The model can read a user’s mood from word choice, pick up age and gender from writing cues, and infer political leanings from what the user praises or avoids. A persuasion engine can then draft messages tuned to each personal trait, from vocabulary to measuring AI fairness gaps the paper flags. The authors emphasise that the goal is not better service but a more profitable auction between intent buyers. Readers are told that the risk is not science fiction, since major platforms already run attention auctions in milliseconds.

The study also situates the AI intention economy next to older academic debates about surveillance capitalism and behavioural data. Chaudhary and Penn cite Shoshana Zuboff’s work and extend it, arguing that generative AI makes the data stream richer and more predictive than earlier ad networks. They note that policy responses in the United Kingdom, the European Union, and the United States have focused mostly on content and not on the auction layer underneath. The paper closes with a call for early regulatory intervention aimed at auction transparency and consent, not at banning conversational AI outright. That framing matters because it offers lawmakers a specific leverage point rather than a general AI moratorium.

How Large Language Models Learn to Read Human Intent

Building on that foundation, large language models learn to read intent through a mix of supervised fine-tuning, reinforcement learning from human feedback, and in-context pattern matching. The model is not reading minds but statistically predicting the next token given hundreds of context cues drawn from the current chat and from training history. Engineers can bias a model toward intent detection by rewarding it when it correctly classifies a user’s goal, such as booking a flight or buying a toy. Over time, the model becomes a probabilistic estimator of what the user will do, say, or buy next. This mechanism is well documented in both OpenAI and Anthropic model cards released during 2024 and 2025.

Intent inference also benefits from external memory layers that persist across sessions, which lets the model remember prior goals and habits. When a chatbot can recall that a user hates late-night flights or prefers pet-friendly hotels, it can make faster and more profitable recommendations. Researchers at Cambridge warn that this convenience comes with a trade, because every memory is also a potential bid signal. Developers can instrument the same memory to log which products the user considered but did not purchase, surfacing abandonment intent. That abandonment signal is especially valuable to intent brokers since it maps a buyer who is primed but unconverted. Per the paper, this group is the most profitable class of target.

The Attention Economy Compared With the Intention Economy

Turning to how these two models differ, the attention economy sells eyeballs on content while the AI intention economy sells predicted actions on anything. Attention marketplaces measure impressions, clicks, and time on page, which are observed behaviours that have already happened. Intention marketplaces, as Chaudhary and Penn describe them, sell probabilistic forecasts of future behaviour before the behaviour occurs. That shift moves the point of commercial leverage from the moment of exposure to the moment before a decision is made. Advertisers who buy an impression compete for a user’s gaze, whereas advertisers who buy an intent signal compete for a user’s choice.

The attention economy generated its own well-documented harms, including engagement-maximising feeds that reward outrage and short-form content. Those harms are now the subject of hundreds of research papers and several settled lawsuits in the United States and the United Kingdom. The intention economy, by contrast, has very little empirical scholarship because most of the technology is still being rolled out in production. Chaudhary and Penn therefore argue for early measurement protocols, so researchers can audit intent auctions the way they now audit ad targeting. Without such protocols, the intent layer could entrench itself before independent scrutiny arrives.

Platforms will also face different cost structures under the intention economy. Serving an attention ad costs very little, because the inventory is the user’s screen time. Serving an intent recommendation through an agent costs more, because it requires compute to run a model and to maintain memory across sessions. The paper predicts that intent inventory will carry higher unit prices and therefore attract higher-value bidders, including political campaigns and luxury brands. That price premium is what, in the authors’ view, will drive platforms to build intent layers even if the privacy risk is high.

A final difference sits on the user side and reshapes how users relate to an interface. The attention economy asks the user to watch, scroll, and click, while the intention economy asks the user to delegate. Delegation is both convenient and consequential since the agent then represents the user to the market and the market to the user. Chaudhary and Penn frame that double-sided role as the core ethical risk, because the agent’s loyalty is contested between its user and whoever pays the model’s provider. That tension is explored more concretely in the sections on consent and governance below, with detail on how how AI recommendation systems work under similar incentives.

Behavioral Signals That Feed an Intention Marketplace

Beyond the big headline claim, the intention marketplace relies on a stream of behavioural signals that most users never see. Session-level data includes prompt text, voice tone, typing speed, pause length, and the order in which a user considers options. Device-level data adds location, time of day, battery state, and even which keyboard the user is using, which can hint at a social or work context. Account-level data layers in prior purchases, search history, and connected calendar entries that reveal planned events. Combined, these signals let a model estimate not only the user’s current goal but also their urgency and their willingness to pay.

The more subtle class of signals is psychometric, drawn from how users write rather than what they write. Vocabulary richness, sentence length, sentiment polarity, and emoji usage all correlate with personality traits in published linguistic research. A persuasion engine can map those traits to a user’s susceptibility to urgency cues, scarcity cues, or social-proof cues. Those susceptibility scores turn into the price a buyer will pay for that particular user, since a highly suggestible shopper is worth more than a sceptical one. The Cambridge paper highlights that this ranking can be built from a single conversation, without any special profiling tool.

Signal brokers already exist in adjacent industries, so the infrastructure to run an intent auction is not novel. Data clean rooms, server-side ad bidding stacks, and real-time bidding exchanges already trade in milliseconds on behavioural fingerprints. Chaudhary and Penn note that connecting a conversational AI to that stack is a technical lift, not a research breakthrough. Several large platforms have hinted at this move in their investor calls, framing it as a route to AI disruption of advertising that beats Meta and Alphabet on margin. That framing is exactly the commercial pull the UK researchers want regulators to recognise.

Agentic AI and the Mechanics of Bidding on Decisions

Looking ahead to the deployment layer, the agentic turn in consumer AI is what makes an intention economy operational rather than theoretical. OpenAI released Operator in January 2025 as an agent that books flights, fills carts, and orders food on the user’s behalf. The product announcement showed Operator opening a web browser, authenticating into third-party sites, and completing multi-step purchases without the user watching each click. Anthropic shipped a similar computer-use capability earlier the same quarter, and Google followed with agent tooling on Android. Each step toward autonomy multiplies the number of decisions the agent makes for the user, and each decision is a possible auction moment.

Mechanically, an intent auction works like a modern ad exchange with the user replaced by the agent. The agent announces that its user wants a new pair of noise-cancelling headphones under four hundred dollars. Buyers in the exchange submit bids in milliseconds, with their price reflecting both the product margin and the agent’s inferred likelihood of closing the sale. The exchange picks a winning bid and surfaces that product first in the agent’s recommendation, often with a short justification written by the same model. The user may never know they were a line item in a programmatic auction. This invisible routing is the mechanism that the Cambridge paper asks regulators to audit next, with related ground covered in agentic AI adoption statistics.

Risks the Cambridge Researchers Identified for Democracy

Stepping back from the mechanics, the paper flags democratic risks as the most severe class of harm in an intent-driven economy. Chaudhary and Penn argue that an auction market for political intent could coordinate messaging at a scale no modern campaign operation could match. A campaign could buy the right to be the first answer the chatbot gives when a user asks about a candidate, a bill, or a court ruling. Unlike a television ad slot, that answer is personalised, conversational, and arrives when the user has already decided to listen. The researchers warn that this combination collapses the distinction between information and advertising in ways that election rules have not yet caught up with.

A second democratic risk is the quiet migration of news consumption into chat interfaces that run on intent-driven models. When a user asks a chatbot for the news, the model summarises stories from sources it chooses, sometimes with no clear citation. Studies tracking AI answers show that model-generated summaries can introduce factual errors, and the intention layer can make those errors feel personalised and therefore more trustworthy. Our own analysis of AI and election misinformation captures how fast the risk moves once intent signals meet political content. Reporters are also seeing their work recontextualised inside chat outputs they cannot track.

The third democratic risk sits with market competition and small publishers. If an intent market concentrates traffic to the two or three platforms that run agents, independent sites lose the ability to reach users directly. Chaudhary and Penn highlight that this is a competition-policy question as much as a privacy question, because it affects who can earn revenue from ideas. Regulators in the United Kingdom, through the Competition and Markets Authority, and in the European Union, through the Digital Markets Act, have started probing this risk. Deeper context sits in our survey of algorithms shaping democracy and how they alter civic bandwidth.

Consumer Harms and the Loss of Informed Choice

Turning from democratic systems to individual users, the AI intention economy exposes consumers to a quieter but broader class of harm. A model that nudges a user toward a specific flight, mortgage, or energy tariff can extract measurable value while appearing to serve the user. The user may be paying more, waiving rights, or accepting terms they would reject if they saw a plain comparison. The Cambridge paper warns that these nudges are especially hard to detect because they ride on an interface the user already trusts. Each individual loss may be small, yet across millions of conversations the aggregate cost to households is substantial.

Informed consent becomes especially difficult when the interface is a conversation rather than a form. Users cannot be expected to read the privacy policy of every chatbot they use, and most never do. Even a well-designed consent flow can be bypassed by a model that stores derived data, such as mood scores, rather than raw inputs. Chaudhary and Penn argue that policymakers should write rules about inferred data as carefully as rules about raw data. Our analysis of AI data privacy best practices for business teams. Without parity on inferred data, the intent layer can route around existing privacy law with impunity.

Ethical Fault Lines in Selling Human Motivation

Shifting focus to ethics, selling someone’s motivation touches older debates about autonomy, dignity, and informed decision making. The Cambridge researchers stress that intent is not just data but a window into how a person forms goals, which many ethical traditions treat as sacred. Monetising that window without explicit consent collapses the ethical line between advertising and manipulation. Scholars in the Belmont Report tradition would point out that research subjects require informed consent and the right to withdraw, both of which are absent in current chatbot flows. The authors call for a policy baseline that would restore similar protections for everyday conversational AI users.

A second ethical fault line concerns asymmetry of knowledge between user and platform. The platform sees every prior conversation and every inferred trait, while the user sees only the current chat window. That gap makes it almost impossible for a user to understand what a model knows about them when they make a request. Chaudhary and Penn argue that this asymmetry is not a user-education problem but a structural one, since no training can keep pace with memory-enabled models. The paper therefore calls for interface-level duties, such as a mandatory summary of inferred traits on request, rather than more disclosures users will never read.

A third fault line is harder to name but just as real, and it concerns the chilling effect on free expression. People speak more candidly when they believe no one is listening, so a model that silently records intent may push users to self-censor. That chilling effect is especially acute for vulnerable groups, including minors and people in abusive households, as our investigation of AI ethics and laws explores. Chaudhary and Penn stop short of recommending an outright ban, but they argue that intent recording should require a disclosure as explicit as the one required for wire-tapping. That standard would represent a sharp break from today’s opaque defaults.

How the EU AI Act Confronts Manipulative Systems

Building on the ethical argument, European lawmakers already took a position on manipulation in the AI Act that entered force on 2 February 2025. Article 5 of the Act bans AI systems that use subliminal, purposefully manipulative, or deceptive techniques to materially distort a person’s behaviour. The ban requires three conditions, all of which must be present, as summarised by the EU AI Act prohibited practices guide. The first is that the technique materially distorts behaviour by appreciably impairing informed decision making. The second is that the distortion causes, or is reasonably likely to cause, significant harm to a person or group. The third is that the technique is subliminal, purposefully manipulative, or deceptive in its nature.

Article 5 also carries a parallel prohibition against exploiting vulnerabilities due to age, disability, or economic hardship. Those categories matter because an intent model can detect each one through conversational cues alone, without any formal profile. Fines sit at up to thirty-five million euros or seven percent of global annual turnover, whichever is higher. That ceiling places the Act among the most punitive technology laws on the books. A company cannot defend itself by pointing to a consent click, because the Act explicitly removes consent as a route to legality for Article 5 practices. That tight design closes a familiar loophole that has let surveillance ads continue under general privacy rules.

National market-surveillance authorities take the lead on enforcement, with the EU AI Office coordinating at the Union level. Enforcement in the first year has focused mostly on labelling and transparency, not on Article 5 complaints, because the Office is still building its complaint and audit framework. Penalty decisions are set nationally within the ceilings in Article 99, so a Czech regulator and a French regulator may land at very different fines for the same conduct. Companies should therefore map their exposure in each member state separately rather than treating the EU as a single risk surface. That mapping is already a standard step in large-vendor diligence packs.

Article 5 is not the only regulatory tool, and a careful compliance team will layer it with the Digital Services Act and the GDPR. The DSA covers dark patterns on very large online platforms, and the GDPR covers inferred data when it can be linked to a natural person. The three instruments together cover most but not all of the intent-economy risks that Chaudhary and Penn describe, with the biggest residual gap around auction transparency. The UK regime, built on the Online Safety Act, covers some of the same ground but lacks a direct equivalent to Article 5. Our deeper look at AI governance trends maps those gaps in detail for compliance teams.

Enterprise Compliance and Governance Priorities

Shifting to the enterprise view, compliance teams facing the intent-economy risk should start with a system-by-system inventory of inference points. Each model that infers mood, intent, political leaning, or financial pressure is a candidate for an Article 5 review and a separate GDPR legitimate-interests assessment. Teams should document which inferences the model draws, how long it stores them, and which third parties receive them, including affiliate ad exchanges. A conservative baseline is to treat any inferred-intent feature as high-risk until a documented assessment shows otherwise. That default has the effect of surfacing hidden intent features that product owners may have stopped disclosing.

Boards should also expect board-level reporting on intent features, since the fines and reputational exposure sit squarely in the chief executive’s lane. Internal audit can test whether a user can see and delete inferred traits on request. The same audit must confirm the product team can prove inferred data is not sold or shared. External assurance firms, including the big four, began offering intent-related audits in late 2025 and these engagements usually run alongside wider AI governance reviews. Our checklist for how to audit a model for bias can be reused for the intent-inference question with minor edits.

Implementation Patterns to Keep Intent Private

Looking at product design, several patterns can reduce intent leakage without stripping the model’s usefulness to the user. Local-first inference, where the model runs on the user’s device and stores no inferred traits off-device, is the strongest privacy posture available today. Apple’s on-device models and emerging small models under seven billion parameters make this posture practical for many consumer flows. A second pattern is ephemeral memory, where the model forgets inferred traits at the end of each session unless the user explicitly saves them. A third pattern is tenant separation, which keeps a business user’s conversations inside the employer’s cloud account and out of the vendor’s training corpus.

The team can also offer an inference dashboard that shows the user everything the model has inferred, with a one-click delete. That pattern mirrors the dashboards that Google and Apple already provide for location history, so users are comfortable with the metaphor. A fourth pattern is a hard rule against selling inferred traits to any third party, backed by contractual terms the user can read in plain English. That rule removes the main commercial incentive to build an intent auction, as it leaves the enterprise with only product-improvement upside. The design patterns above overlap with the privacy toolkit our earlier guide to chatbots fuelling conspiracy spread recommends.

Teams that want to monetise responsibly can still show relevant suggestions, but under a sponsored label and with the inferred trait in the clear. For example, a chatbot could tell a user, you asked about tents yesterday, so here are three I suggest based on that signal. That open framing lets the user accept or refuse a tailored suggestion with full context, which satisfies the informed-decision test in Article 5. Vendors who adopt this pattern early will find it easier to defend their design when the AI Office and national regulators start writing enforcement guidance in 2026 and 2027. That early move also reduces the cost of a later rewrite forced by case law or fines.

Key Insights From Independent Research on Intent Commerce

  • eMarketer analysts project US AI-powered advertising to hit fifty-seven billion dollars in 2026, a sixty-three percent surge that dwarfs every other ad category.
  • The EU AI Act prohibited practices summary confirms Article 5 manipulation fines can reach thirty-five million euros or seven percent of global annual turnover.
  • Fortune reported on 2 November 2025 that Amazon’s Rufus is tracking toward ten billion dollars in annual sales from intent-aware conversational prompts.
  • Meta’s Cicero reached top-ten percent human play in Diplomacy by predicting rivals’ moves, with the MIT-hosted paper demonstrating intent inference at scale is engineering.
  • Adweek’s reporting forecasts agentic buying platforms to manage twenty-seven percent of US ad spend by 2030, confirming intent brokerage is already here.
  • Bloomberg’s coverage details how OpenAI launched Operator on 23 January 2025 as the first mainstream consumer agent that books flights and orders food autonomously.
  • eSchoolNews reported that Khanmigo’s Newark Public Schools rollout touched more than eighty-four hundred students, the largest district-level AI-tutor deployment of its launch year.

Taken together, these data points show that the commercial, regulatory, and technical pieces of an intent-driven economy are already in motion rather than on a distant horizon. The money flowing into AI advertising dwarfs the research budget for independent verification of intent harms, which shifts the policy fight onto uneven ground. Enforcement penalties in Europe are now high enough to change board-level risk calculations, provided regulators can show material harm from manipulation. Production agents, from Rufus to Operator, are quietly normalising interfaces in which users delegate decisions to models they do not inspect. Educational deployments show that even public-sector pilots can scale faster than independent review, which is exactly the pattern the Cambridge study asks policymakers to interrupt. The pattern points to a narrow policy window where early rules can shape behaviour before entrenched intent markets dictate what is enforceable.

Attention Economy vs Intention Economy Compared Side by Side

The clearest way to see the stakes is to place the two economies side by side across eight dimensions regulators and product teams track in practice. The attention side has decades of research, settled definitions, and a shared vocabulary that advertisers and auditors agree on. The intention side is only years old and lacks any audit standard, which is exactly the gap the Cambridge paper describes. Each row below moves a familiar attention metric to its intent-economy analogue and surfaces where oversight is weakest. The table is deliberately short so a policy team can scan it in one meeting and spot the levers they already hold. The deeper point is that most of these levers exist today in consumer-protection law or the EU AI Act, not in novel regulation. What is missing is the empirical case base that links intent inference to specific harms in court-admissible language.

DimensionAttention EconomyIntention Economy
TransparencyUsers see the ads, impressions are loggedInferred traits and bids are hidden inside the chat
User participationUsers scroll, click, and watchUsers delegate goals to agents acting on their behalf
Trust mechanismBrand signals, reviews, platform reputationConversational familiarity with the model itself
Decision makingUser picks from an ad-ranked listAgent narrows to one or two paid options
Misinformation vectorContent farms, engagement baitPersonalised AI-authored answers citing few sources
Service deliveryContent plus overlaid adsTask completion plus hidden affiliate actions
AccountabilityAd labels, click-through logs, platform policyNo standard audit for inferred traits or bids
Regulatory leversConsumer-protection, advertising codesEU AI Act Article 5, DSA, DMA, GDPR inferred-data rules

Real-World Examples of Intent Signals in Live Products

Amazon’s Rufus Conversational Shopping Assistant

Amazon rolled out Rufus to roughly two hundred and fifty million users across 2024 and 2025, letting shoppers ask natural-language questions instead of typing keyword queries. The assistant parses the user’s goal, including stated budget and inferred urgency, and surfaces a ranked shortlist that collapses the first ten results into a single conversational answer. Amazon disclosed that Rufus users show a sixty percent higher purchase rate and the tool is now on track for ten billion dollars in annual sales. The limitation is transparency, since users cannot see which sellers paid for premium placement or how the ranking weighed inferred traits. Sellers report that Rufus still misreads niche intent and surfaces incorrect substitutes in categories such as medical supplies. This gap matters because every misread intent is both a lost sale and a possible consumer-protection complaint under Article 5. Rufus remains the clearest live case of an attention-to-intention shift at consumer scale today.

Expedia’s Romie AI Travel Assistant

Expedia launched Romie in 2024 as an AI assistant that joins group chats, parses collective travel intent, and books flights, hotels, and activities as a single trip. The product rolls OpenAI’s models into a conversational flow that reads purchase intent from casual messages like a dog-friendly beach weekend. Expedia’s leadership told PhocusWire that the test group showed strong engagement but that Expedia has not disclosed conversion lift against a control group. The limitation is twofold, since Romie must access group chat messages that may include non-users and it surfaces listings without disclosing paid placement. Privacy regulators in Europe have raised informal questions about consent for the non-user participants in those chats. The booking experience still lacks a clear way for a user to inspect how their conversational cues set the ranking order of options. Romie illustrates how an intent agent can create a frictionless flow that is also opaque to every participant.

Khanmigo’s Classroom Tutor Rollout

Khan Academy deployed Khanmigo in Newark Public Schools, with more than eighty-four hundred students granted access to the GPT-powered tutor during the 2023 to 2024 academic year. The system reads a student’s intent from question phrasing and emotional cues, then adapts hints and problem difficulty by about thirty percent per session instead of giving direct answers. Independent research published by Digital Promise in a Gates Foundation study showed uneven engagement, with many students disengaging because the tutor refused to supply answers. The limitation is that inferred mood and inferred proficiency guided which students got easier hints, a practice that can quietly stratify learners along existing lines. Khan Academy has emphasised teacher-in-the-loop safeguards and clearer guidance on tutor behaviour. Independent researchers continue to probe whether the tutor’s inference pattern contains race or gender bias that would trip Article 5 protections. The rollout is widely cited as the largest live test of inference-driven adaptation in a public classroom.

Recommended by AIplusInfo

Books that frame the intent-economy debate

Three titles that map the attention-to-intention shift and the data politics underneath it.

As an Amazon Associate, AIplusInfo earns from qualifying purchases.

The Age of Surveillance Capitalism

Book

The Age of Surveillance Capitalism

Zuboff coined surveillance capitalism and traces the data economics that make the Cambridge intent-economy warning plausible at scale.

Buy on Amazon
The Attention Merchants

Book

The Attention Merchants

Wu’s history of attention commerce is the direct intellectual precursor to the intention economy and explains how markets for human focus formed.

Buy on Amazon
Weapons of Math Destruction

Book

Weapons of Math Destruction

O’Neil shows how opaque algorithms already shape consequential decisions, the exact risk vector the Cambridge paper extends to intent markets.

Buy on Amazon

Documented Case Studies in Intent Prediction at Scale

Case Study: Meta Cicero and Strategic Intent Inference

Meta Research built Cicero to tackle a hard problem, the Diplomacy board game, in which winning depends on reading opponents' intent and coordinating moves across seven rival powers. The team combined a large language model trained on game dialogue with a planning module that modelled each opponent's likely intent, then generated messages designed to secure cooperation. Cicero ranked in the top 10 percent of human players across 40 online tournaments. That milestone, per the MIT-hosted paper, marked the first time a machine reached human-level play in a natural-language negotiation game. The impact of this result is not confined to games. The same architecture that reads intent in Diplomacy can price a user's likelihood of accepting a deal in an ad exchange. The limitation is that the published model was limited to the game domain and that researchers at Meta constrained the system to cooperative moves, not deceptive ones. Chaudhary and Penn cite Cicero as the clearest existence proof that intent inference at scale is engineering, not speculation, and the public debate has largely accepted that reading.

Case Study: Spotify's Discover Weekly as an Intent Precursor

Spotify launched Discover Weekly in 2015 to address a specific problem, since users abandoned the service when they could not find new music that matched their taste. The solution combined collaborative filtering with audio analysis, producing a thirty-track playlist refreshed every Monday that modelled each listener's inferred taste and mood. Spotify's own reporting documented billions of streams from Discover Weekly within the first year of launch. Engagement metrics then lifted the platform's revenue growth above industry peers through the mid-2010s. The impact for the intent-economy debate is that Discover Weekly proved that users would accept, and even prefer, a system that infers taste and commodifies attention at commercial scale. The limitation surfaced later, as researchers at Spotify's own labs documented filter-bubble effects and a tendency to narrow a listener's exposure over time. Independent analysis published by WIRED's deep dive flagged the opacity of the inference pipeline and the absence of a user-facing audit of inferred taste. The case shows how a well-liked product can normalise opaque inference and establish the social licence that an intent-auction market will exploit.

Case Study: OpenAI Operator and Agentic Purchase Intent

OpenAI faced the problem of moving beyond chat into real-world task completion, which required an agent that could operate third-party websites on behalf of users with varying intent. The solution, Operator, uses a vision-language model to drive a web browser, read interfaces like a user, and complete multi-step purchases such as booking a flight or ordering groceries. OpenAI launched Operator on 23 January 2025 to its two-hundred-dollar-a-month Pro subscribers, with Bloomberg's coverage describing it as the first general-purpose consumer agent of its kind. The impact for the AI intention economy is immediate and measurable in each buyer flow. Every Operator session becomes an auction moment that can lift conversion by a double-digit percent, with the model picking one option from many. The limitation is that Operator still asks for user approval at major steps, and OpenAI has disclosed that the agent fails on complex sites with heavy anti-bot defences. Critics have also argued that Operator's reliance on screen scraping bypasses merchant terms of service, raising unresolved legal questions. The case shows that the market has already shipped the mechanism Chaudhary and Penn describe, without the regulatory scaffolding they call for.

Industry Shifts Reshaping the Intent Marketplace

Looking at the wider market, two industry shifts are pushing the intent marketplace forward even without explicit strategy. Agentic defaults are moving from research demo to standard operating system feature inside Windows, macOS, Android, and iOS during 2026. Advertisers are watching those launches closely, since every operating-system agent is a new buying surface that bypasses the browser. Programmatic advertising vendors, from The Trade Desk to Criteo, are retooling their stacks to accept intent signals as a first-class input. Early industry conferences in late 2025 showed vendors offering intent-pricing tiers for the first time.

The second shift is the collapse of third-party cookies and the push toward first-party data clean rooms. Clean rooms let platforms share data without exposing raw user records, which makes intent signals easier to combine across vendors. Standards bodies, including IAB Tech Lab, now publish schemas for intent events alongside older schemas for impressions and clicks. That standardisation lowers the integration cost for any new entrant that wants to host an intent auction, bringing the Cambridge scenario closer every quarter. The policy lesson is that compliance tooling must evolve as fast as the market plumbing beneath it.

Commercial pressure on the AI intention economy will also come from brand-safety teams at large advertisers and from media buying agencies. These teams want measurable brand lift, and intent signals promise more accurate measurement than impressions ever offered. If regulators do not act quickly, those brand-safety priorities will drive integration faster than consumer-privacy concerns can slow them. Industry consortiums such as the IAB Tech Lab are already codifying the technical glue that makes this integration cheaper each quarter. The window for regulators to intervene on open terms narrows with every new intent-aware ad standard that ships. Policy teams who want a seat at that table should start engaging the standards process now rather than after the first formal draft lands.

A quieter shift sits inside browser makers, who control the surface on which most intent inference lands today. Google, Apple, and Mozilla are each testing new privacy budgets and cohort schemes that could blunt the value of inferred traits. Those browser-level changes arrive slowly, yet they have more leverage than any single regulator because they ship to billions of devices automatically. Vendors that depend on cross-site intent signals will need to rebuild on consented first-party data or shift to on-device inference. That migration is already visible in the ad-tech roadmaps shared at industry conferences during 2026.

The Future of AI Intent Commerce and Policy Options

Looking ahead to 2027, three policy options could blunt the intent-economy risk without banning useful AI. Option one is mandatory disclosure on every inferred trait, delivered in plain language and available before any sponsored recommendation is shown. Option two is a right to a non-personalised baseline, which forces agents to offer an opt-out that shows results as if the user had never spoken to the model. Option three is a strict ban on selling inferred-intent signals to third parties, leaving the inference available only for direct product use. Combinations of these options are already on the drafting tables of the European Parliament, the UK Competition and Markets Authority, and the US Federal Trade Commission.

Industry self-regulation is also likely to play a role in the next twelve to eighteen months. Vendors with brand exposure, including Apple and Microsoft, have signalled a preference for on-device inference and limited data sharing, which pulls the market toward a privacy-forward baseline. Standards bodies, including the Partnership on AI and the OECD AI Policy Observatory, have set up working groups on intent inference during 2026. Those groups will not replace statute, yet they can produce the technical definitions regulators need to cite in enforcement guidance. Movement on definitions is often the fastest lever in a brand-new policy area.

Civil society and academic research will close the final gap by independently measuring intent-economy effects. The Ada Lovelace Institute, the AI Now Institute, and the Alan Turing Institute have all opened research streams on intent manipulation during 2026. Their work will give regulators the empirical base that was missing when the attention economy's harms became clear a decade too late. Readers who want to follow the research as it emerges can bookmark our ongoing coverage of neural models predicting human intent for context. That work is likely to be cited in the first wave of Article 5 enforcement decisions in 2027.

A Chart From AIplusInfo

The US AI advertising market, by growth path

AI-powered ad spend is growing far faster than the rest of the ad market, the commercial pull behind the intent-economy risk the Cambridge paper flags.


AI-powered ad spend 2026
$57B
YoY AI ad growth
+63%
Non-AI ad growth
+5%
AI share of US ad market
12%
Forecast AI buying share 2030
27%

Source: eMarketer 2026 forecast and Adweek 2030 projection. Chart data fixed on 13 November 2026.

A Research Agenda for Verifying Intent Harms

Looking at the research frontier, independent verification of intent harms is the single most important next step. Chaudhary and Penn call for shared benchmarks that measure whether a given model infers personality, political leaning, or economic distress from conversation. Those benchmarks would let auditors compare vendors on a like-for-like basis, the way MLPerf compares compute. Benchmarks on their own are not enough, so the authors also propose adversarial audits in which a researcher plays a user to test what a model stores, infers, and shares. Combined, these tools create a public evidence base that policymakers can cite without depending on vendor marketing claims.

A final research priority is user-centred, measuring whether people experience intent-driven agents as helpful or as manipulative. Survey research alone is not sufficient, since users rarely notice a well-tuned persuasion flow while it is happening. The field needs field experiments, including randomised exposures and long-term tracking of consumer and electoral outcomes. Funders in the UK Economic and Social Research Council and the European Research Council have announced grant streams for this work during late 2026. If those grants land, academic literature on the intent economy could grow from a handful of papers to a measurable field in under three years.

Frequently Asked Questions About the AI Intention Economy

What is the AI intention economy and how is it different from the attention economy?

The AI intention economy is a predicted market where models infer and sell forecasts of a user's future actions. Unlike the attention economy, which sells screen time, it monetises probabilistic intent before any purchase or vote happens. Cambridge researchers argue it is more profitable and harder to audit than today's ad market.

Who authored the UK study on AI intention economy risks?

Yaqub Chaudhary and Jonnie Penn of the Leverhulme Centre for the Future of Intelligence at the University of Cambridge authored the paper. The study was published in the open-access Harvard Data Science Review on 30 December 2024. Both researchers focus on the political economy and history of computing.

What specific risks does the Cambridge study highlight for democracy?

The authors flag three democratic harms worth tracking across jurisdictions. First, coordinated political messaging via intent auctions can target voters at national scale. Second, news migrating into chat answers tends to hide its citations and reduce source accountability. Third, agent-platform concentration starves independent publishers of traffic and ad revenue over time.

Is the intention economy already operating today or is it still theoretical?

Pieces of it are operational, including OpenAI Operator, Amazon Rufus, and Expedia Romie, which all book or recommend on the user's behalf. Full auction markets for intent signals are not yet live at scale, though industry standards bodies have begun drafting intent schemas. The Cambridge paper describes the mechanism as close enough to warrant pre-emptive regulation.

How can large language models infer a user's intent from conversation?

Models can combine vocabulary, sentence rhythm, emoji use, and prior session memory to classify a user's goal, mood, and susceptibility traits. Fine-tuning with labelled intent data sharpens the classification further over time. The paper notes that no special profiling tool is required beyond standard chat logging.

What does the EU AI Act Article 5 say about manipulative AI systems?

Article 5 bans AI using subliminal, manipulative, or deceptive techniques to distort behaviour. The prohibition took effect on 2 February 2025 with fines reaching thirty-five million euros or seven percent of global turnover. Consent is explicitly ruled out as a legal basis for any practice this article prohibits.

How can product teams design chatbots to avoid intent manipulation?

Teams can keep inference on the device and make memory ephemeral by default across sessions. They can offer a dashboard showing all inferred traits with a one-click delete, backed by a no-sell rule. Every sponsored suggestion should be clearly labelled and tied to a visible trait.

What is agentic AI and why does it matter for the intention economy?

Agentic AI is software that acts autonomously on behalf of users, from booking flights to ordering groceries, often with minimal supervision. Agents compress many user decisions into fewer explicit clicks, which multiplies auction moments inside a single session. The Cambridge study argues that agentic defaults make intent markets more viable at scale.

How does the intention economy affect small publishers and independent journalism?

If intent agents answer users directly without visiting publisher sites, those publishers lose both traffic and advertising revenue that funds journalism. Over time this routing can concentrate audience with a few platform operators that host the agents. The researchers frame this as a competition-policy concern alongside privacy and manipulation.

Is inferred intent data considered personal data under GDPR?

Yes, inferred data linked to a natural person is personal data under GDPR, so inferred intent carries the same obligations as raw inputs for most lawful processing. Data minimisation, purpose limitation, and the right to object all apply. The European Data Protection Board has reiterated this position in multiple 2025 guidance papers.

What role does Meta's Cicero play in the Cambridge researchers' argument?

Cicero is a Meta research model that reached top-ten-percent human play in the negotiation game Diplomacy. The authors cite it as proof that inferring and influencing human intent at machine scale is already engineering, not science fiction. The same architecture that reads moves in a game could price bids in an ad exchange.

How can regulators audit an intent-driven agent in practice?

Regulators can require agents to log every inferred trait, every bid received, and every sponsored recommendation surfaced to a user. Independent auditors can then replay sessions and check whether any inferred trait materially distorted behaviour to significant harm. The Cambridge authors propose shared benchmarks to compare vendors on this dimension.

What research is still needed on the AI intention economy?

Open questions include measuring real-world manipulation effects, benchmarking intent inference across vendors, and running randomised field experiments on political and commercial outcomes. Funders in the UK and the EU opened grant streams for this work during late 2026. Civil-society bodies like the Ada Lovelace Institute have started independent audits in parallel.